On September 11, OpenAI announced it will retire custom GPTs across every ChatGPT plan. Its migration FAQ, updated last week, puts the scheduled retirement at December 11, 2026. On that date custom GPTs and their pages become inaccessible. That gives a local business 74 days to rescue whatever it built on top of them. If your quote writer, your review-reply bot, or your new-hire SOP assistant lives inside a custom GPT, it is now on a countdown. The migration tool will not move everything, and the parts it drops are usually the parts that made the GPT useful.
What actually retires on December 11
Start with the facts, because the coverage is messy. Custom GPTs are the build-your-own assistants OpenAI launched in November 2023. The replacement is plugins, which bundle reusable instructions (now called skills), reference files, and connected apps. OpenAI’s FAQ says the change affects all ChatGPT plans. Personal accounts on Free, Go, Plus, and Pro already lost the ability to create or publish new GPTs, though existing ones still run. Enterprise workspaces get a creation freeze on October 26 and a retirement on December 11, with a February 11, 2027 extension only for workspaces with an approved deferral.
One wrinkle catches small businesses off guard. What matters is where a GPT was built, not which plan you pay for. If your office manager relies on a public GPT made by a software vendor or a franchise group, that GPT follows its creator’s timeline. It could go dark on December 11 even if you never built anything yourself. Access to the old GPT also does not guarantee access to its replacement plugin. Somebody on your team has to confirm the new version works for them before the old one disappears.
What survives the move and what breaks
OpenAI built a one-click path. In My GPTs, the creator selects Migrate to plugin. The GPT’s instructions become a skill. Its knowledge files get copied into the plugin’s reference files. Connected apps come along as apps. For a simple GPT that holds your price sheet and a page of instructions, that covers most of it. Migration uses the latest published version only, so drafts and unpublished edits do not transfer. After you migrate, the original GPT goes read-only, and you cannot delete it. Finish your edits first.
Here is what does not survive. Custom actions, the API calls that let a GPT look up a job in your CRM or push a lead into your scheduling tool, do not transfer at all. The selected model does not carry over. Sharing settings reset, so the plugin starts private and nobody on your team can use it until you share it again. Past conversations stay behind. OpenAI also warns that a migrated plugin may respond differently than the GPT did. That last line matters most for anything customer-facing, like a GPT that drafts estimates or replies to reviews in your voice.
A four-step rescue you can run this week
Block two hours on Tuesday morning. That is enough for a shop running three to five GPTs. Pull whoever uses them most into the room, because they know which outputs matter and which prompts break. The goal is not a perfect migration by Friday. The goal is to know exactly what you have, save the parts only you control, and have a test set ready so you can judge the replacement on facts instead of a gut feeling.
- Inventory. Open a shared sheet with five columns: GPT name, who built it, who uses it, the job it does, and whether it calls a custom action. Include GPTs your team uses that someone else built. Most shops find one or two they forgot about, usually a pricing helper or an email template bot a former employee set up.
- Export the business, not the bot. For every GPT you own, copy the full instruction text into a plain document. Download every knowledge file. Store both in a folder you control, like a shared drive. Your instructions and your price sheets are the asset. ChatGPT was only the container holding them.
- Build a test set. For each GPT, save five real inputs and the output you accepted last time. Example: five customer quote requests and the five estimates you sent. Add one hard case, like a job with an unusual scope or an angry review. This is how you will grade the replacement.
- Migrate the easy ones first. GPTs with no custom actions can move now, if the option shows in your account. Run your test set through the new plugin. If four of five outputs match what you would have sent, share it with the team and move on. If not, tighten the skill instructions and rerun the set.
A typical case shows how this plays out. Picture a two-location dog grooming business with a GPT, built in 2024, that turns a customer’s booking notes into a pre-appointment text. It holds a pricing PDF, a breed-by-breed time guide, and one page of tone rules. That GPT has no custom actions, so it migrates in about ten minutes. The work is the test set. Five past bookings, five texts that went out, and one hard case where a customer asked about a matted coat. Twenty minutes of grading tells you whether the plugin sounds like your shop.
Where to rebuild a GPT that used custom actions
GPTs that called an external service are the real project. OpenAI’s own guidance says rebuilding that connection may require a custom MCP server and technical setup. For a 10-person business with no developer, that is a hard sell. You have three practical routes, and you should pick based on how often the integration runs and how much it costs you when it breaks.
First, check whether an app in the ChatGPT plugin directory already covers the job. OpenAI replaced its App Directory with a Plugin Directory on July 9. Search it for your CRM, booking tool, or invoicing software by name. If your custom action only read contacts or created a deal, a prebuilt app may do it with no code. Test it against the same five inputs before you trust it with live customer data.
Second, move the integration out of ChatGPT entirely. A Zapier or Make scenario you already pay for can take a form submission, call an AI model, and write to your CRM without depending on any chat product’s feature roadmap. Third, if the action was a convenience and not a core workflow, drop it. Paste the lookup result into the chat by hand for a few weeks and see if anyone misses the automation.
The lesson here goes past OpenAI. Any vendor can retire the container your process lives in. This summer alone, ChatGPT swapped its default models several times, rebuilt its app directory, and retired the official DALL-E GPT. The businesses that shrug these off keep their instructions, examples, and reference files in a document they own, then point whichever tool is current at that document. Treat your prompt library like your customer list. It belongs on your drive, not only inside a vendor’s product.
The signal to watch before Halloween
Watch for the in-product migration notice on your own account over the next 30 days. OpenAI targeted the migration experience for Enterprise workspaces on September 22, with a second banner wave targeted for October 1. Timing for Plus, Pro, and Business accounts arrives through in-product notices, not a published calendar. Check the My GPTs page every Monday, and have whoever built each GPT do the same in their own account, since the notice follows the creator.
When the Migrate to plugin button appears, your inventory and test sets from this week turn a scramble into a one-hour job. If the button still has not appeared by October 26, when Enterprise creation ends, stop waiting and rebuild your most critical GPT as a plugin by hand from your exported instructions. Seventy-four days sounds like plenty. For most local businesses, the holiday rush will eat a big share of it. Do the inventory now, while the calendar is still quiet, and put a December 1 reminder on the calendar to confirm every replacement works.
